A Graph Neural Network Charge Model Targeting Accurate Electrostatic Properties of Organic Molecules.

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Machine Learning in Chemistry

Background:

• Traditional partial charge assignment methods (e.g., RESP, AM1-BCC) are computationally expensive and conformer-dependent.
• Existing Graph Neural Network (GNN) models often reproduce AM1-BCC charges, which approximate higher-level calculations.

Purpose of the Study:

• To investigate the suitability of various charge assignment schemes (ESP, AIM) as training targets for GNN-based charge models.
• To develop new GNN charge models that combine the strengths of different approaches for improved condensed phase modeling.

Main Methods:

• Investigated ESP and Atoms-In-Molecule (AIM) based schemes as training targets for GNNs.
• Co-trained GNN models using AIM charges, molecular dipoles, and electrostatic potentials.
• Collected quantum mechanical AIM properties at high-level theory (ωB97X-D/def2-tzvpp) in vacuum and implicit solvent.
• Trained new GNN charge models and scaled charges between vacuum and solvated sets.

Main Results:

• Demonstrated that co-training GNNs with AIM charges and electrostatic properties is effective.
• Developed new GNN charge models trained on high-level quantum mechanical data.
• Showcased that charges can be scaled between vacuum and solvated sets for force field development.
• Integrated GNN charges with optimized Lennard-Jones parameters for polarized condensed phase force fields.

Conclusions:

• The developed GNN charge models offer a fast and flexible alternative for partial charge assignment.
• The approach enables the creation of accurate, polarized force fields for condensed phase simulations.
• Applied the charge models to study electrostatics-driven structure-activity relationships in medicinal chemistry.